Methods for image recovery in computational imaging

As a classic topic that has been studied for decades, image restoration is still a very active research area. Developing more effective and efficient methods are highly desirable. This thesis addresses image restoration problems for applications in computational imaging including Time-of-Flight (ToF) imaging and digital photography. While ToF cameras have shown great promise at low-cost depth imaging, they suffer from limited depth-of-field and low spatial-resolution. We develop a computational method to remove lens blur and increase image resolution of off-the-shelf ToF cameras. The method solves latent images directly from the raw sensor data as an inverse problem, and supports for future ToF cameras that use multiple frequencies, phases and exposures. Photographs taken by hand-held cameras are likely to suffer from blur caused by camera shake during exposure. Removing such blur and recovering sharp images as a post-process is therefore critical. We develop a blind deblurring method that is purely based on stochastic random-walk optimization. This simple framework in combination with different priors produces comparable results to the much more complex state-of-the-art deblurring algorithms. Blur causes even more serious issues for document photographs as slight blur can make Optical Character Recognition (OCR) techniques fail. We address the blind deblurring problem specifically for common document photographs. Observing that the latter are mostly composed of high-order structures, our method captures such domain property by a series of high-order filters as well as customized response functions. These parameters are trained from data by discriminative learning approach and form an end-to-end network that can efficiently and jointly estimate blur kernels and legible images. Discriminative learning approaches achieve convincing trade-off between image quality and computational efficiency, however, they require separate training for each restoration task and problem condition, making it time-consuming and difficult to encompass all tasks and conditions during training. We combine discriminative learning and formal optimization techniques to learn image priors that require a single-pass training and share across various tasks and conditions while keeping the efficiency as previous discriminative methods. After being trained, our method can be combined with other likelihood or priors to address unseen restoration tasks or further improve the image quality.Science, Faculty ofComputer Science, Department ofGraduat